Fix grammar JSON spacing bug

Fix structural regressions

Fix: 10 critical NLP logic bugs in grammar, spelling, and punctuation to prevent false positives

Fix plural adjective agreement regression for broken plurals

Fix 30 NLP edge cases in Grammar, Spelling, and Punctuation (Phase 10 results and Extension UI improvements)

fix(grammar): add hamza fixes for ايضا, اول, and position-aware إنّ/أنّ rules

update benchmark logs

Fix collision benchmark failures: patch overlap mapping, tokenization guard, and rule hallucinations

Fix punctuation alignment and hallucinations, add punctuation safety rules, add kana/inna grammar rules

feat(nlp): implement case-checking rules for kana and inna sisters using camel_tools

Add debugging scripts and fix grammar/punctuation logic

fix: verb suffix stripping, regex unnormalized hamzas, and POS-independent triggers to reach 70%

refactor: simplify architecture and replace whitelists with CamelTools

Fix: Enhance Jazm and Nasb logic with camel_tools for Af'al Khamsa and defective verbs

FIX-38/39 + Layer 1/2/3: Benchmark normalization + spelling safety + grammar rules

FIX-33b: Protect second+third preposition regexes with blocklist

FIX-33/34/35: Fix infinite loop — grammar corruption + spelling conflict

Fix grammar rules output log — remove hardcoded ... truncation marker

Phase 14c: Fix FIX-31 whitespace bug + fix misleading log truncation

Fix: TerminalPunctuationGuard for short phrases and comprehensive Nasb/Jazm grammar rules fixes

Feat: Add demonstrative and noun-adjective dual/plural agreement rules

Fix: Prevent conditional sentences rule from overcorrecting 'إن' and 'من' by requiring the subsequent word to be a verb

Fix: Preserve input punctuation during grammar correction

HOTFIX: Remove broad preposition/nasb regex rules — caused Hallucination 0%, +38 FP

Critical: Isolate grammar rules in try/except so regex fallback ALWAYS runs

Phase 13: Grammar SV/gender whitelist, preposition/nasb regex, pronoun-h guard

Phase 12: Fix 6 batches — grammar pipeline bypass, religious/structured protection, punct rearrangement, SV/gender agreement

Phase 11: Master Bug Fix — 23 fixes across 8 files

Phase 8: Fix all critical bugs from adversarial validation

fix(critical): stop spelling from corrupting correct words + fix pronoun agreement + reject hallucinations

fix: all model bugs — S1 S2 S3 G1 P1 (6 fixes across 4 files)

fix: Grammar Gradio client retry with exponential backoff (3 attempts, 2s/4s/8s)

fix: Grammar retry on rate-limit + cursor position after correction

fix: Grammar rules false positive - الامتحان was incorrectly matched - Increased minimum stem length from 2 to 4 in fix_prepositions_advanced - Prevents root-level ان endings (الامتحان, الإنسان) from being treated as grammatical ون/ان

feat: NLP-2 Grammar Integration — dependency-aware pipeline - Add grammar_rules.py (ArabicGrammarGuard from Grammer_Rules.py with camel-tools) - Add grammar_service.py (lazy-loaded Gradio Client + rules pipeline) - Wire /api/grammar endpoint to real grammar service - Wire /api/analyze Step 2 to grammar (receives AraSpell output) - Update /api/health to report grammar status - Add gradio_client to requirements.txt - Add camel_data download to Dockerfile - Frontend unchanged (already supports grammar-error yellow highlighting)

feat: Professional editor UI redesign - 3-column layout (docs | editor | sidebar) - Documents panel always visible on desktop - Import/export moved to left sidebar - Removed CTA button from nav - Replaced all emojis with SVG icons - Clean toolbar with separator - Mobile: docs stay as overlay sidebar

fix: Remove 'Apply All' button - conflicts with spelling+grammar pipeline

fix: Word-level grammar suggestions + robust grammar model validation

fix: Prevent overlapping spelling+grammar suggestions from duplicating text - Backend: Skip grammar suggestions that overlap with existing spelling ranges - Frontend: applyAllSuggestions skips overlapping ranges to prevent duplication

revert: Restore ORIGINAL grammar params that were working (num_beams=4, max_length=512) - Remove all hallucination checks that were blocking valid corrections - The model WAS correcting ذهب→ذهبت before, only highlight was wrong - Keep word-level diff fix (in app.py) + raw output logging

fix: Switch grammar to greedy decoding + stronger anti-hallucination - num_beams=1 (greedy, most stable for seq2seq correction) - repetition_penalty=2.0 (strong anti-repeat) - Log raw model output before hallucination check - Tightened bigram duplicate detection

fix: Prevent grammar model hallucination (rearranging/duplicating text) - Dynamic max_length (1.3x input instead of fixed 512) - repetition_penalty=1.5 to discourage repeating - Reduce beams 4→2 (less hallucination on CPU) - Add _is_hallucinated() safety checker: output length, word count, repeated phrases, sentence rearrangement

feat: NLP-2 Grammar model integration (Bayan_Arabic_Grammar T5) - New: src/nlp/grammar/grammar_service.py (lazy T5 loading) - Dockerfile: pre-download grammar model (2.14GB safetensors) - Wire /api/analyze Step 2 to grammar service - Update /api/grammar standalone endpoint - Add _grammar_available() health check